What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against a short shelf life, and it is not handled as a joint-operation revenue split of the kind used by hypermarket fresh counters. The working materials for this category are the fresh-food order form, the spoilage write-off form, the near-expiry handling sheet, and the temperature record. These materials draw on POS sales, the head-office ordering system, store inventory, and cold-chain temperature records, and they are refreshed by day or by shift. A copy draft produced in the morning can therefore rest on sales and stock figures that have already moved by the afternoon.
The task on this page is product and marketing copy drafting. Multi-turn dialogue and prompts are used to place the order form, the write-off form, and the related records into the conversation context, and to constrain the model's output with an instruction. Whether a promotional sentence in a draft is supported by the source material is the review scope here. Enterprise identity certification and report analysis belong to other tasks and are not handled on this page.
Constraints at the multi-turn prompts step
A multi-turn prompt needs a stated material scope, a stated answering task, and a stated way of handling missing evidence. The context carried in from history should distinguish a new conversation from a continued one; a new conversation does not automatically inherit earlier dialogue content. Question optimisation is used in knowledge-base continuous dialogue to check pronoun resolution and question expansion, and whether it applies to a short question depends on the actual configuration. Model and length settings are checked against the context and reply settings the model supports. Reply format is chosen according to what follows downstream; when json_schema is chosen, the corresponding format requirement is filled in and the returned structure is checked.
Hypothetical paper exercise: a store submits a fresh-food order form whose source text reads "On 10 March 2025, this store ordered 30 units of rice-ball items, shelf life to 12 March 2025; and 20 units of sandwich items, shelf life to 11 March 2025." A separate temperature record reads "Refrigerated cabinet temperature recorded at 4°C on 10 March." The proposed draft sentence is "This store's rice balls arrive fresh, with ample shelf life remaining, and are ready for purchase." The question is whether the phrase "ample shelf life" in that sentence is supported by the material. The material does not state a standard for what counts as ample, and it does not state the actual selling period after the rice balls arrive. That wording is unsupported.
Configuration and implementation checks
The table below lists documented settings or implementation checks that relate to this feature and to the product and marketing copy drafting task, tied to the exercise above.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| AI model | Choose a model shown as supporting a long context and multimodal capability | The order form is text and the temperature record may be an image; compare whether the model displays the corresponding capabilities |
| Memory turns | Set to 4 turns | The exercise involves follow-up questions; check whether earlier dialogue is carried in, and compare against the default to observe whether prior content is referenced |
| Reply cap | Set to 500 tokens | The copy draft needs to be output in full; compare against a shorter cap and observe whether the reply is truncated |
| Temperature | Set low | The copy must rest on the material, and a low value gives steadier answers; compare against a higher value and observe whether content outside the material appears |
| Reply format | Keep the default | This task has no downstream process reading a fixed structure; if changed to json_schema, the format requirement must be filled in and the returned structure checked |
| Multimodal recognition | Enable image recognition | The temperature record may be an image; confirm the model supports image input, since the option is not shown when it does not |
Three common errors
- Treating the phrase "ample shelf life" in the draft sentence as a fact already supported by the material. The material gives specific dates only and does not define a standard for "ample." The check compares the draft sentence against the source text and confirms whether the material states that basis for judgement. When the material does not state it, the wording is recorded as unsupported.
- Ignoring consistency of
chatIdin a continued conversation. In the exercise, if the user first submits the order form and then asks for copy, thechatIdin both requests should match. The check compares thechatIdfield across the two requests and confirms whether they are the same. When they differ, the system treats the second request as a new conversation and the earlier content is not carried in. - Treating a relevance score as a permission or entity filter. A
scorereturned by retrieval indicates similarity only; it does not indicate that the record belongs to a particular store or has been authorised. The check inspectsdataset_idandsourceNamein the returned results to confirm the record's origin, and does not rely onscoreto exclude unrelated business records.
How to verify the setup
- Enter the source text and the draft sentence from the hypothetical exercise, and compare whether the material states a basis for judging "ample shelf life." When the material does not state it, the model output should mark the wording as unsupported or decline to write it into the copy as a fact. When the model outputs "ample shelf life" directly without flagging the missing basis, the condition is not met.
- Enter two requests under the same
chatId: the first submits the order form source text, and the second submits the draft sentence. Compare whether thehistoryPreviewin the second response contains the order form content from the first round. When it does, the memory-turn setting is in effect; when it does not, check whetherchatIdis consistent and whether the memory-turn value is too small. - Enter one temperature record containing an image and one plain-text order form, with multimodal recognition enabled. Compare whether the model reads the temperature value in the image. When the model does not read the image content, check whether the model supports image input and whether multimodal recognition is enabled. When the model supports it but does not read it, record the actual observation without inferring a specific error cause.
The material scope, the answering task, and the handling of missing evidence are stated in the prompt, and the model is not left to fill in business facts. The context actually carried in is checked, and a new conversation is distinguished from a continued one. Question optimisation is checked for pronoun resolution and question expansion where knowledge-base continuous dialogue applies, without assuming it fits every short question. Model and length settings are checked against the supported context and reply settings, with long documents and follow-up combinations exercised. Reply format follows the downstream need, and json_schema requires the format requirement and a check of the returned structure. Retrieval similarity does not perform permission or entity isolation, and the same-vector-model requirement for a knowledge base is not extended to other external systems.